🚀 Nous Research ships token-stacking, pre-training 2.5× faster
Audio in Mandarin Chinese · English transcript below
⚡ US-China chips in a choke-hold, Google dukes it out with Apple’s whole stack, training now 2.5× faster, open-source + BNPL all ship—AI is twisting SaaS into an octopus on the rush-hour subway!
Global AI competition is accelerating across the board. This acceleration is fueled not only by significant technical breakthroughs, such as Nous Research's method of improving LLM pre-training efficiency by 2.5 times through token stacking, but also by intensified competition stemming from US restrictions on chip sales, which in turn prompts China to accelerate domestic substitution and launch models like DeepSeek V4. Amidst this backdrop, the in-depth discussion on Hacker News regarding "teaching Claude to understand causality" precisely highlights that while the industry pursues speed and scale, AI's deep understanding and alignment remain the core challenges determining its future value.
Today's Top 3 Headlines
- AI Industry News
🤖 Googlebook debuts this fall, Gemini Intelligence across Android suite
Google pre-I/O: first Googlebook this fall from Acer/ASUS/Dell/HP/Lenovo; Gemini Intelligence hits Galaxy & Pixel this summer, then watches, cars, glasses. AI-native hardware gives devs fresh edge.
Source ↗ - AI Industry News
🤖 AI data centers gulp 6% of UK/US power; 3 GW global “zombie” waste
IDC: UK & US data centers each draw 6% of national power—2× global avg; US "zombie" apps idle 3 GW. AI teams face power queues; efficiency + clean energy decide survival.
Source ↗ - AI Technology
🤖 Nous Research token-stacking pretrain 2.5× faster
Nous Research drops Token Superposition Training, cutting 270M-10B-param pretraining 2.5×. Cash-strapped startups can now iterate faster, cheaper, and divert saved GPU budgets to data & post-train finesse.
Source ↗
+8 more headlines
- 🤖 Klarna BNPL lands in Google Search & Gemini; US Google Pay adds installments
- 🤖 Hackers ask: how to make Claude grasp the real “why”
- 🤖 YouTube test: DeepSeek doubles coding speed, new AI-code paradigm
- 🤖 US chip curb 2.0, China counters with Deep Seek V4
- 🤖 Wu Minghui: AI Kills SaaS, 1800-staff firm open-sources multi-Agent net "Octopus"
- 🚀 AMD Ryzen PRO 9000 debuts 3D V-Cache for biz
- 🤖 DeepSeek open-sources LLM, costs plunge—sector rally ahead
- 🤖 FFmpeg/VLC creator on Lex: decoding libs power Chrome & 100M TVs